One-Class Classification Using ℓp-Norm Multiple Kernel Fisher Null Approach.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Science

Background:

• The one-class classification (OCC) problem involves identifying outliers or novelties within a dataset, assuming only normal data is available for training.
• Traditional methods often struggle with high-dimensional data and the need for robust feature representation.

Purpose of the Study:

• To propose a novel one-class multiple kernel learning (MKL) algorithm for enhanced OCC.
• To develop an efficient optimization strategy for the proposed MKL framework.
• To explore concurrent learning of related OCC tasks through shared kernel weights.

Main Methods:

• Development of a one-class MKL algorithm based on the Fisher null-space OCC principle.
• Incorporation of $\ell _{p}$-norm regularization for kernel weight learning.
• Formulation of the problem as a min-max saddle point Lagrangian optimization task with an efficient solver.
• Extension for concurrent learning of multiple related OCC tasks.

Main Results:

• The proposed one-class MKL approach demonstrates superior performance compared to baseline and other algorithms.
• Effective optimization achieved through the proposed min-max saddle point strategy.
• The concurrent learning extension shows promise for related OCC problems.

Conclusions:

• The presented one-class MKL method offers a powerful and efficient solution for the OCC problem.
• The approach is validated across diverse datasets, highlighting its generalizability.
• Future work can explore further extensions and applications of this MKL framework.